Wean Hall Slam Experiments : 07.08.2004

Experiment of July 8th, 2004 : Wean Hall 6th floor

Conducted by Deryck Morales (deryck@cmu.edu)

Experiment Motivation:

These experiments confirm autonomous global localization behavior given the problems of similar (even identical) edge feature maps and mistakenly driving through strong meet points.

Experiment Setup:

Both of these experiments took place in an environment that is a subset of the 6th floor of Wean Hall here at Carnegie Mellon. I used cardboard to seal the hallways. The corresponding graph is shown to the left. The Atlas for the environment was previously created by the robot using the user oracle mechanism for loop closure. The features for edges E₀ and E₂ were arranged so as to produce identical feature maps.

Experiment A Results:

The robot was placed in the hallway near node zero (N₀). After accessing edge E₀ and arriving at node zero, the robot loaded the Atlas for the environment. Comparing its observed location and the Atlas, the robot estimated that it had arrived at either node zero (N₀) or node three (N₃).

The robot then traversed edge E₀ and arrived at node one (N₁).

Comparing the observed node information, the robot estimated that it had arrived at either node 1 or node 2. The observed edge feature map was a match for both E₀ and E₂, and so the robot was able to narrow down it's location to either node 1 from edge 0, or node 2 from edge 2.

The robot then "turned right" and traversed edge E₁, arriving at node two (N₂).

Comparing the observed node to the Atlas, the robot estimated that it had arrived at either node 1 or node 2. The observed edge feature map was only a match for E₁ from node 1 to node 2, and the robot successfully localized itself as being on N₂ arriving on E₁ with 99.99% probability.

It is worth noting that without the edge feature map, the robot still would have successfully localized itself using the topological turn test, with a slightly lower probability. (still well above 90%).

Experiment B Results:

The robot was placed in the hallway near node three (N₃). After accessing edge E₂ and arriving at node three, the robot loaded the Atlas for the environment. Comparing its observed location and the Atlas, the robot estimated that it had arrived at either node zero (N₀) or node three (N₃).

The robot then traversed E₂ and drove through N₂, because I was standing in the corridor of E₁, blocking that passage off to simulate error due to sonar noise. The width of the E₁ hallway is very close to our limit for valid pathways, and so sensor noise may cause that passageway to appear "closed off".

The robot then traversed E₃ and drove through N₁, this time due to local sonar map noise "closing off" the passage to E₁, as mentioned above.

The robot then continued across edge E₀, arriving at node zero (N₀).

Comparing the observed node and traversed edge to the Atlas, the robot found that no edge feature map matched its observation, and so did not apply those test results. The edge traversal distance eliminated any single edge traversal hypothesis, and supported a few missed meet point accounts. Specifically, the robot estimated that it had either began at node 0, traversed E₀,E₃,and E₂, and arrived at node 3, or that it had taken the reverse of this path. Therefore, the robot could only estimate that it was either at node 0 or node 3.

The robot traversed E₀ and arrived at node 1 (N₁).

The robot's observations narrowed down its possible location to either node 1 from edge 0, or node 2 from edge 2.

The robot traversed E₁ and arrived at node 2 (N₂).

Finally, the robot's observations allowed it to estimate its location on node 2 from edge 1 with 99.99% certainty. As with exeriment A above, it is worth noting that the robot would have successfully localized itself at this point with the topological turn test alone, but with slightly less certainty (still well above 90%).

Discussion:

These experiments were performed in a relatively small environment (33% of the total size of Wean Hall 6th floor), but the demonstrated process scales to cases where there are sets of identical edge feature maps and symmetric subsets of the environment. The critical tests for localization in these environments are the topological turn test and the edge feature map test. The edge feature map test is strong because one unique edge feature map is sufficient to determine which edge was traversed as well as the direction of the traversal. The topological turn test is strong because it captures consecutive edge connectivities, or the "shape" of sequential edge pairs.

The next step in these tests would be to map environments with symmetry and identical edge feature configurations on a larger scale.

Go back to: H-SLAM Experiments Page | H-SLAM page

Conducted by Deryck Morales (deryck@cmu.edu)

Experiment Motivation:
These experiments confirm autonomous global localization behavior given the problems of similar (even identical) edge feature maps and mistakenly driving through strong meet points.

Experiment Setup:
	Both of these experiments took place in an environment that is a subset of the 6th floor of Wean Hall here at Carnegie Mellon. I used cardboard to seal the hallways. The corresponding graph is shown to the left. The Atlas for the environment was previously created by the robot using the user oracle mechanism for loop closure. The features for edges E₀ and E₂ were arranged so as to produce identical feature maps.

Experiment A Results:
	The robot was placed in the hallway near node zero (N₀). After accessing edge E₀ and arriving at node zero, the robot loaded the Atlas for the environment. Comparing its observed location and the Atlas, the robot estimated that it had arrived at either node zero (N₀) or node three (N₃).

	The robot then traversed edge E₀ and arrived at node one (N₁).

	Comparing the observed node information, the robot estimated that it had arrived at either node 1 or node 2. The observed edge feature map was a match for both E₀ and E₂, and so the robot was able to narrow down it's location to either node 1 from edge 0, or node 2 from edge 2.

	The robot then "turned right" and traversed edge E₁, arriving at node two (N₂).

	Comparing the observed node to the Atlas, the robot estimated that it had arrived at either node 1 or node 2. The observed edge feature map was only a match for E₁ from node 1 to node 2, and the robot successfully localized itself as being on N₂ arriving on E₁ with 99.99% probability. It is worth noting that without the edge feature map, the robot still would have successfully localized itself using the topological turn test, with a slightly lower probability. (still well above 90%).


Experiment B Results:
	The robot was placed in the hallway near node three (N₃). After accessing edge E₂ and arriving at node three, the robot loaded the Atlas for the environment. Comparing its observed location and the Atlas, the robot estimated that it had arrived at either node zero (N₀) or node three (N₃).

	The robot then traversed E₂ and drove through N₂, because I was standing in the corridor of E₁, blocking that passage off to simulate error due to sonar noise. The width of the E₁ hallway is very close to our limit for valid pathways, and so sensor noise may cause that passageway to appear "closed off".

	The robot then traversed E₃ and drove through N₁, this time due to local sonar map noise "closing off" the passage to E₁, as mentioned above.

	The robot then continued across edge E₀, arriving at node zero (N₀).

	Comparing the observed node and traversed edge to the Atlas, the robot found that no edge feature map matched its observation, and so did not apply those test results. The edge traversal distance eliminated any single edge traversal hypothesis, and supported a few missed meet point accounts. Specifically, the robot estimated that it had either began at node 0, traversed E₀,E₃,and E₂, and arrived at node 3, or that it had taken the reverse of this path. Therefore, the robot could only estimate that it was either at node 0 or node 3.

	The robot traversed E₀ and arrived at node 1 (N₁).

	The robot's observations narrowed down its possible location to either node 1 from edge 0, or node 2 from edge 2.

	The robot traversed E₁ and arrived at node 2 (N₂).

	Finally, the robot's observations allowed it to estimate its location on node 2 from edge 1 with 99.99% certainty. As with exeriment A above, it is worth noting that the robot would have successfully localized itself at this point with the topological turn test alone, but with slightly less certainty (still well above 90%).



Discussion:
These experiments were performed in a relatively small environment (33% of the total size of Wean Hall 6th floor), but the demonstrated process scales to cases where there are sets of identical edge feature maps and symmetric subsets of the environment. The critical tests for localization in these environments are the topological turn test and the edge feature map test. The edge feature map test is strong because one unique edge feature map is sufficient to determine which edge was traversed as well as the direction of the traversal. The topological turn test is strong because it captures consecutive edge connectivities, or the "shape" of sequential edge pairs. The next step in these tests would be to map environments with symmetry and identical edge feature configurations on a larger scale.